Deep learning-based classification network for glaucoma in retinal images

Abstract

Glaucoma is the second leading cause of blindness, resulting in damage of the optic nerve. Ophthalmologists perform diagnosis by a retinal inspection of widened pupils. Machine learning approaches, until now being guided and arduous, demand automated solutions. As a result, deep networks owing to self-learning can provide automated diagnostic procedures in less time. This study presents Classification of glaucoma network (CoG-NET), which is a deep network for prediction of glaucoma. Based on the experimental results, 93.5% Accuracy, 0.95 Sensitivity, and 0.99 Specificity are observed for the proposed network, which outperforms commonly used State-of-the-art. It can be observed that the proposed model gives an Area under receiver operating characteristics curve (AUROC) of 0.99. Further, corresponding feature activation maps for the proposed network validate its focus centralized to the optical disc and cup exclusively present in the retinal fundus image for glaucoma diagnosis that can be employed for glaucoma screening at the initial stages.